Financial network communities and methodological insights: a case study for Borsa Istanbul Sustainability Index

The paper situates financial networks as a framework to understand market dynamics by mapping relationships among stocks and sectors via measures like correlations, volumes, and return dependencies. Such networks reveal communities (clusters) and influential nodes that relate to systemic risk and contagion. The authors highlight the growing relevance of ESG factors (environmental impact, social practices, and governance quality) in investment decisions and portfolio resilience. They identify a gap: traditional financial network analyses rarely integrate ESG metrics, potentially missing sustainability-driven interdependencies and clustering. The study’s objective is to integrate ESG scores into multiple network models (correlation, continuous/discrete mutual information, and linear/nonlinear causality) for firms in the Borsa Istanbul Sustainability Index (XUSRD), detect communities via Leading Eigenvector and Girvan-Newman algorithms, and test whether ESG sub-scores influence community formation using nonparametric methods (Kruskal-Wallis, Conover, log-rank). The expected contribution is a more comprehensive understanding of how sustainability practices shape market connectivity and community structures, informing sustainable investment strategies.

The review synthesizes evidence linking ESG performance to financial outcomes and market stability. Studies report generally positive associations between ESG and profitability, credit quality, stability, and portfolio performance (Kim & Li 2021; Sinha et al. 2019; Aybars et al. 2019; Caporale et al. 2022; Luo et al. 2024; Park & Lee 2023; Park & Oh 2022). ESG integration in investment processes can support value and risk management but faces measurement challenges (Chan et al. 2020; Zaccone & Pedrini 2020; Ma 2023). ESG also contributes to financial stability by mitigating systemic risk (Stolbov & Shchepeleva 2022; Lupu et al. 2022; Ling et al. 2023). Evidence on market value impacts varies across sectors and geographies (Ionescu et al. 2019; Lee & Isa 2023; Yoon et al. 2024; Junius et al. 2020). The review includes related macroeconomic work by Batrancea et al., emphasizing governance, social infrastructure, and environmental rules in growth. It also highlights ESG’s growing role in predictive analytics and machine learning for markets and ratings (Ang et al. 2023; Giese et al. 2019; Lee et al. 2024; Gong et al. 2024; Raza et al. 2022; Zhang & Zhang 2023). The review motivates integrating ESG into network analysis to identify systemic risks and enhance model predictive power.

Data: 78 companies in the Borsa Istanbul Sustainability Index (XUSRD); daily closing prices with 249 observations per firm. Two were omitted due to missing/duplicate codes. ESG scores (2022/2023 Refinitiv-based) include sub-scores: Emissions, Resource Use, Innovation, Human Rights, Product Responsibility, Workforce, Community, Management, Shareholders, CSR Strategy. Network construction: Nodes are companies; edges are weighted according to each approach and filtered via Triangulated Maximally Filtered Graph (TMFG). 1) Correlation network: Compute log returns and Pearson correlations; define correlation distance dc(i,j)=√2(1−ρij). Lower distances indicate stronger links; TMFG filters to 3V−6 edges. 2) Mutual information networks: - Continuous MI (Kraskov estimator) on log returns; - Discrete bivariate MI after binning returns. Edge weights are reciprocals (1/MI), emphasizing weaker MI with higher weight; TMFG filtering applied using these weights. 3) Causality networks: - Linear causality via Granger tests (VAR with lag order p=1 specified in results), assign edge weights using p-values (smaller p→lower weight/stronger link), TMFG filtering; - Nonlinear causality via Transfer Entropy significance tests, p-value-based weights and TMFG filtering. 4) Community detection: Leading Eigenvector (spectral modularity) and Girvan-Newman (iterative removal by weighted edge betweenness), optimizing weighted modularity. 5) Comparative analyses: Community comparison across methods using metrics: Normalized Mutual Information (NMI), Variation of Information (VOI), Split-Join Distance (SJD), Rand Index (RI), Adjusted RI (ARI). Novel random-walk-based comparison: For networks with weights where lower implies stronger tie, define transition probabilities Pij ∝ 1/wij and analyze visit distributions (10,000 steps). Compute transfer entropy between visit-frequency series across networks (lags=1, quantile threshold=0.1, Shannon entropy, 100 shuffles, 300 bootstraps). Statistical assessment of ESG influence on communities: ESG sub-scores tested across community memberships per network and community algorithm using nonparametric tests (Kruskal–Wallis, Conover post-hoc, and log-rank/Gehan). Additionally, linear regression relates ESG overall scores to sub-scores, and ANOVA decomposes contributions.

• Dataset and regression diagnostics: Adjusted R-squared=0.9402; R-squared=0.948; AIC=377.12; BIC=405.44, indicating strong explanatory power of ESG sub-scores for overall ESG score. Regression coefficients (Estimate, p-value): Resource Use 0.117 (6.7×10⁻⁶), Innovation 0.054 (6.6×10⁻⁶), Human Rights 0.050 (0.039), Product Responsibility 0.089 (0.00014), Workforce 0.173 (6.7×10⁻⁶), Community 0.140435 (1.5×10⁻⁸), Management 0.197 (1.3×10⁻²¹), Shareholders 0.041 (0.00117), CSR Strategy 0.036 (0.018876); Emission 0.0279 (0.188). ANOVA shows strong effects across sub-scores (e.g., Resource Use F=418.749, p=1.5×10⁻³⁰; Emission F=224.437, p=4.5×10⁻²³; Management F=194.79, p=1.6×10⁻²¹, etc.). - Network structures: Across correlation, mutual information (continuous/discrete), and causality (linear/nonlinear) TMFG-filtered networks, clear sectoral clustering is observed, especially within financial and manufacturing sectors. Mixed-sector communities also appear, indicating inter-sector dependencies. - Community detection outcomes: Both Leading Eigenvector and Girvan-Newman reveal sectoral clusters. In correlation networks, manufacturing and finance firms form tight-knit communities. In continuous and discrete MI networks, strong informational linkages produce both intra-sector and cross-sector communities; central firms (e.g., DOAS, MAVI, BIMAS, TUPRS) have high weighted degrees and eigenvector centrality. In linear causality networks, sectoral causal pathways are pronounced, with financial hubs (e.g., AKBNK, GARAN) and manufacturing leaders (TOASO, VESTL, ARCLK) central. Nonlinear causality networks reveal additional cross-sector causal links and more complex clusters involving energy, manufacturing, finance (e.g., ZOREN, SISE, TOASO; GARAN, EREGL). - Community comparison metrics (Table 5): Linear causality network performs best in recovering coherent communities: VOI=0.866 (lowest among most), NMI=0.7593 (highest), SJD=32 (lowest), RI=0.899, ARI=0.6356. Correlation network shows moderate performance (NMI=0.5883, ARI=0.4722). Continuous MI performs weakest (NMI=0.3039, ARI=0.1977). Discrete MI moderate (NMI=0.4528, ARI=0.422). Nonlinear causality shows good performance (VOI=0.573, NMI=0.573, RI=0.8151, ARI=0.3471) but with higher SJD. - Random-walk transfer entropy between networks (Table 4): Notable directed information flows include Nonlinear→Linear causality TE=0.113118 (largest), Linear→Nonlinear=0.039678, Discrete MI→Linear=0.0280271, Correlation→Nonlinear=0.028385. These indicate intertwined linear and nonlinear dynamics and that correlation/MI structures inform causality networks. - ESG influence on community formation (selected significant results): • Correlation network: Leading Eigenvector—Conover shows Resource Use significant (p=0.0167); log-rank highlights Product Responsibility (p=0.0425). Girvan-Newman—Kruskal-Wallis Innovation (p=0.0265) and CSR Strategy (p=0.008); log-rank Innovation (p=0.012), CSR Strategy (p=0.005). • Continuous MI: Girvan-Newman—Kruskal-Wallis Emission (p=0.032); log-rank Emission (p=0.034), Innovation (p=0.021). • Discrete MI: Leading Eigenvector—Kruskal-Wallis and log-rank CSR Strategy significant (p≈0.0028–0.0031); Conover Resource Use (p=0.0154). Girvan-Newman—Conover Community (p=0.029). • Linear causality: Leading Eigenvector—Human Rights significant across tests (KW p=0.012; Conover p=0.003; log-rank p=0.022). Girvan-Newman—log-rank Innovation (p=0.02), Emission (p=0.034). • Nonlinear causality: Leading Eigenvector—KW Innovation (p=0.019), Human Rights (p=0.020); log-rank Emission (~0.056), Resource Use (p=0.045), Innovation (p=0.022), Human Rights (p=0.038). Girvan-Newman—log-rank Human Rights (p=0.023), Shareholders (p=0.059); Conover Product Responsibility (p=0.013), Workforce (p=0.079). Overall, ESG sub-scores—especially Emission, CSR Strategy, Innovation, and Human Rights—significantly shape community structures across network types, with causality-based networks yielding the clearest community recoveries.

The study examines whether ESG performance explains how firms cluster within different financial networks. Nonparametric tests across community partitions show that several ESG sub-scores significantly differ across communities, implying that firms with similar sustainability practices tend to co-locate in the same clusters. The importance of particular ESG dimensions varies by network type and community algorithm: Resource Use, Innovation, CSR Strategy, and Product Responsibility bear on correlation and MI communities; Human Rights strongly influences linear-causality communities; and Emission, Resource Use, Innovation, and Human Rights feature in nonlinear-causality communities. These patterns indicate that sustainability practices correspond to both contemporaneous comovement and directional influence structures in markets. Community comparison metrics further suggest that causality-based networks better capture meaningful partitions in the XUSRD universe, consistent with financial linkages that propagate shocks. Random-walk transfer entropy reveals substantial directed information flows between linear and nonlinear causality networks, and from correlation/MI structures to causality, highlighting multi-faceted dependencies. Together, the findings support integrating ESG metrics into network analysis for improved understanding of market connectivity and for informing sustainable investment decisions.

The paper integrates ESG metrics into multiple financial network models for firms in the Borsa Istanbul Sustainability Index and evaluates community structures via two algorithms. Results show robust sectoral clustering—especially in financial and manufacturing sectors—along with mixed-sector communities that reflect inter-sector dependencies. Community comparison metrics favor causality-based networks in terms of partition quality. Crucially, ESG sub-scores—particularly Emission, CSR Strategy, Innovation, and Human Rights—significantly influence community formation across several networks, implying that sustainability practices contribute to cohesive market communities and may enhance portfolio resilience. Contributions include: (i) a multi-model network analysis with TMFG filtering; (ii) dual community detection and cross-method comparisons; (iii) a random-walk-based comparative framework using transfer entropy; (iv) empirical evidence that ESG metrics shape financial community structures. Future research should: incorporate intraday data and more markets; conduct longitudinal (dynamic) network analyses; integrate alternative ESG sources (e.g., third-party ratings, alternative data); employ adaptive/machine learning models for time-varying networks; and study regulatory impacts and cross-market comparisons. Developing real-time network-ESG monitoring tools could aid investors and policymakers in fostering resilient, sustainable markets.

Key limitations include: reliance on daily closing prices (omitting intraday dynamics); ESG data limited to publicly available sources (potential measurement/disclosure biases); static network snapshots that do not capture temporal evolution of relationships; focus on the XUSRD universe limits generalizability to other markets; and nonparametric tests, while robust, may not capture complex interacting effects among ESG factors and network structures.